Graph Deep Learning for Long Range Forecasting

Abstract

<p>Deep learning-based models have been recently shown to be competitive with, or even outperform, state-of-the-art long range forecasting models, such as for projecting the El Niño-Southern Oscillation (ENSO). However, current deep learning models are based on convolutional neural networks which are difficult to interpret and can fail to model large-scale dependencies, such as teleconnections, that are particularly important for long range projections. Hence, we propose to explicitly model large-scale dependencies with Graph Neural Networks (GNN) to enhance explainability and improve the predictive skill of long lead time forecasts.</p><p>In preliminary experiments focusing on ENSO, our GNN model outperforms previous state-of-the-art machine learning based systems for forecasts up to 6 months ahead. The explicit modeling of information flow via edges makes our model more explainable, and it is indeed shown to learn a sensible graph structure from scratch that correlates with the ENSO anomaly pattern for a given number of lead months.</p><p> </p>

Spatial-temporal transformer network for multi-year ENSO prediction

“…As deep learning techniques have developed, researchers have began to design neural networks for predicting weather elements (e.g., rainfall), which can well mine complex and intrinsic correlations, such as artificial neural networks (ANN) Feng et al (2016) 2022), CNN-LSTM Zhou and Zhang (2022), graph neural networks (GNN) Cachay et al (2020), recurrent neural network (RNN) Zhao et al (2022), transformer Ye et al (2021a) etc. Feng et al Feng et al (2016 propose two methods to predict the existence of ENSO, and the time evolution of ENSO scalar features, which provided a new prediction direction for predicting the occurrence for ENSO events.…”

“…Mu et al Mu et al (2019) defined ENSO prediction as a spatio-temporal series prediction issue and used a mixture of ConvLSTM and rolling mechanism to predict the outcome over a longer range of events. The GNN was first used in Cachay et al (2020) for seasonal prediction, it predict the result in a longer lead time. Zhao et al Zhao et al (2022) designed an endto-end network, named Spatio-Temporal Semantic Network (STSNet), it provided a multiscale receptive domains across spatial and temporal dimensions.…”